Fast and incremental loop closure detection with deep features and proximity graphs

In recent years, the robotics community has extensively examined methods concerning place recognition task within scope of simultaneous localization and mapping applications. This article proposes an appearance-based loop closure detection pipeline named “Fast Incremental Loop Detection (FILD++). First, system is fed by consecutive images and, via passing them twice through a single convolutional neural network, global local deep features are extracted. Subsequently, hierarchical navigable small-world graph incrementally constructs visual database representing robot's traversed path based on computed features. Finally, query image, grabbed each time step, set to retrieve similar locations route. An image-to-image pairing follows, which exploits evaluate spatial information. Thus, in proposed article, we propose network for feature extraction contrast our previous work (FILD), while exhaustive search verification process adopted over generated avoiding utilization hash codes. Exhaustive experiments eleven publicly available data sets exhibit system's high performance (achieving highest recall score eight them) low execution times (22.05 ms average New College, largest one containing 52,480 images) compared other state-of-the-art approaches.